The present invention relates to data center infrastructure, and more particularly, this invention relates to utilizing a low-latency lossless switch fabric in a data center.
Low latency is a highly desirable feature for data center switch fabric. For example, in high-frequency transactions, low latency allows applications to execute large volumes of orders, such as automated stock trades, etc., at fractions of a second. Similarly, in real-time communications, such as video feeds, telemetry, etc., delays in processing information may be detrimental to user experience or efficient control of devices relying on the video feeds and/or telemetry.
An important problem for low latency switch fabric implementations is that they do not provide deep buffering, and hence packets are lost when the fabric is congested. That is, a switch is not capable of forwarding a packet due to congestion conditions and the switch drops one or more packets, which causes a failure or significant delay of the transaction.
Existing solutions for lossless switches involve internal packet buffering. A buffered switch is configured to send all packets through a memory buffer to avoid packet loss. Unfortunately, this solution causes increases in latency because moving a packet into and then out of memory takes time, thus increasing latency for the solution. Accordingly, a better solution would be beneficial to provide a low-latency lossless switch fabric in a data center.